Why do banks dare to lend based on token usage volume?

Because usage volume is a more real-time and harder-to-manipulate operating metric than financial reports: it is recorded by third-party model service platforms, continuous, fine-grained, and checkable daily, directly reflecting the real user usage intensity of the enterprise's product. The cost of fraud is far higher than polishing statements. For asset-light, collateral-poor AI startups, this is credit creation "using real usage volume instead of collateral"; for banks, it opens up a new customer base that traditional financial frameworks cannot reach.

Where are the risks of Token Loans?

There are three main risks: First, usage volume can be brushed; the moral hazard of enterprises colluding with model vendors to brush volume objectively exists. Second, usage volume fluctuates greatly; model upgrades or user loss could cause the credit basis to collapse instantly. Third, there is no mature market for the disposal of collateral. Currently, products are generally small-scale pilots with whitelist systems, and banks are supporting this with cross-verification of data. It looks more like the beginning of a new dimension for risk pricing rather than a mature financial product.

Banks Start Lending to AI Companies Based on 'Tokens': Why Usage Volume Can Serve as Collateral

Q: What are "Token Loans"?

Token Loans are credit products offered by banks to AI enterprises. The core basis for credit assessment shifts from fixed assets and financial reports to token usage volume—the real API call data of enterprise large models. On September 13, 2026, the first batch of "Yiqi Token Loans" landed in Beijing Economic-Technological Development Area, with 6 banks approving nearly 2 billion yuan in credit to AI industrial chain companies. The People's Daily published a full-page in-depth report the next day. Guangzhou Haizhu District launched Guangdong's first special financial product for the token economy at the end of August.

Hotspot Tracking: Hotspot Release × Technical Judgment × Practical Advice. Author: Yongliang

title: “Hotspot Tracking 005 | Banks Start Lending to AI Companies Based on ‘Tokens’: Why Usage Volume Can Serve as Collateral” cover: cover.png author: Yongliang digest: ""

Hotspot Tracking: Hotspot Release × Technical Judgment × Practical Advice. Author: Yongliang

On September 13, the first batch of “Yiqi Token Loans” landed in Beijing Economic-Technological Development Area (E-Town), with 6 banks approving nearly 2 billion yuan in credit to AI industrial chain enterprises in the area. A day later, People’s Daily published a full-page investigation in its “Finance In-Depth” section, asking directly: What do “Token Loans” bring to AI enterprises? Looking at the longer timeline: At the end of August, Guangzhou Haizhu District released Guangdong’s first special financial product for the “Token Economy”; in early September, Beijing E-Town followed suit; at the China International Fair for Trade in Services (CIFTIS), multiple banks put “Computing Power Loans” and “Computing Power Insurance” on display, with ICBC showcasing “Computing Power Infrastructure Loans” and “Computing Power R&D Loans.” Within a week, “lending to AI companies based on token usage” went from a local pilot to a national topic. This is not ordinary financial news—behind it lies the AI industry placing “real usage volume” in the position of collateral for the first time.

What Happened

Let’s start with the product itself. “Token” (Ciyuan) is the Chinese name for the smallest unit of information processed by large models: every time you ask AI to write a copy or run a round of inference, the bill is calculated in tokens. In March this year, the National Data Bureau designated the Chinese name for Token as “Ciyuan,” positioning it as the “value anchor of the intelligent age.” The growth in volume lives up to this title—the national daily average token usage volume was 100 billion in early 2024, jumped to 100 trillion by the end of 2025, and exceeded 140 trillion this March, growing more than a thousandfold in two years.

The pain point “Token Loans” aims to solve is very specific: The most valuable assets of AI enterprises are algorithms, models, and data, but these don’t count in the eyes of banks—light assets, weak financial reports, scattered cash flows, no factories, and lack of collateral, they are naturally blocked from traditional credit. A section chief from the Haizhu District Investment Promotion Bureau calls this a “mismatch”: banks value physical assets, while these companies’ assets are code and orders. So, banks everywhere changed the “ruler” for credit granting: they directly turn operational traces like monthly token consumption data, computing power service contracts, and API call logs into credit credentials to issue pure credit loans.

The speed of implementation is faster than imagined. Guangzhou Haizhu District took the lead in releasing “Token Loans,” with Bank of China, CITIC Bank, and Guangzhou Bank participating in the layout: Guangdong’s first loan went to Tengyuan Digital, which does digital marketing, a 3 million yuan, 3-year pure credit loan approved by the Bank of China Guangzhou Haizhu Branch; an AI short drama company obtained 5 million yuan based on daily computing power consumption, and another AI e-commerce live streaming company got 3 million yuan based on computing power purchase orders. In Shenzhen Longgang, ICBC approved 5 million yuan for Aizhihui Technology, a specialized and sophisticated “Little Giant” enterprise doing embodied intelligence training, via “Computing Power e-Loan.” China Construction Bank Guangdong Branch, in collaboration with the Guangdong Token Trading and Service Center, launched “Token Loans,” approving 30 million yuan for an intelligent IoT company in Nansha. The scale in Beijing is larger: the first batch of “Yiqi Token Loans” in E-Town approved nearly 2 billion yuan in credit; National Intelligent Mobility Control (NIM), which does autonomous driving computing platforms, received a 30 million yuan credit line from CITIC Bank, and Shenzhou Everbright, which does computing power infrastructure, received a 30 million yuan credit limit expansion from Industrial Bank. Policy openings have also appeared simultaneously: this July, the People’s Bank of China and nine other departments jointly issued a notice supporting the extension of credit evaluation from “looking at collateral” to “looking at data.”

What Does This “New Ruler” Measure?

It is worth stopping to think clearly: On what basis can token usage volume serve as proof of credit?

The key lies in two points: real-time, and hard to manipulate.

First, real-time. Traditional credit looks at financial statements, which are compiled quarterly or annually—by the time the data comes out, the business heat has long cooled off for three months to a year. Token usage volume is a running account: how many orders the company took today, how much volume the model ran, it can be reflected in the interface logs readable by the bank within the month. An AI company’s revenue might not have picked up yet, but usage volume has already started moving; conversely, once business stops, usage volume drops that same month. For AI companies where “brand projects have long payment cycles, downstream payments lag, but upstream computing power procurement must be prepaid,” this time difference is exactly the source of cash flow breaks—usage volume happens to be the indicator closest to the operational site.

Second, hard to manipulate. Financial statements can be polished, and cash flows can be made to look pretty; but API call logs and computing power usage records are automatically generated by machines, naturally tamper-proof, and banks can verify them by accessing the backend of computing power platforms. Local officials involved in product design explained this logic: cleaning and aggregating massive raw logs creates a visualized operational report—credit officers who don’t understand code can understand the curves. Moreover, usage volume is hard to fake: to make the curve look good, you have to spend real money to tune models and burn computing power, and the cost is real. Banks then cross-verify with computing power procurement contracts, downstream service contracts, and technical qualifications, and a “computing power bill” becomes a near real-time projection of the company’s operating status.

A detail reveals the granularity of this logic: the head of an AI short drama company in Guangzhou calculated that the token cost per minute for ordinary quality short dramas is about 1,500 yuan, while high quality can reach 4,000 yuan—based on daily consumption data, this company applied for and received a 5 million yuan, 3-year credit loan. This is also why product forms in various places are highly convergent: Bank of China Guangzhou Branch divides customers into supply, application, and service categories, with a maximum credit of 30 million yuan per customer; China Construction Bank looks at upstream contracts and accounts receivable, downstream consumption and orders, and intermediate platform settlement flows; Beijing E-Town’s ruler is wider—besides Token consumption, it also looks at technical barriers, team strength, and the degree of independent controllability of core technologies. The names are varied, but the ruler is the same: using real usage volume instead of collateral.

The Calm Half

While watching the excitement, one must watch the risks even more. The People’s Daily investigation wrote very bluntly about industry concerns: there have long been chaos in the market such as reselling idle tokens and tampering with consumption speeds in the backend; if banks lend based on manipulated data, the assessment will be distorted. People from the credit departments of several banks also admitted: simply taking token data as the sole basis might encounter invalid calls, brushing orders to inflate consumption, or artificially inflating business scale—the threshold for fraud isn’t that much higher, it’s just burning money to brush volume, and the money for brushing volume itself can come from loans.

Token prices themselves can also fluctuate. They follow large model call demand and computing power market trends. Taking something with a floating price as a reference for credit granting, will the assessment become distorted when the market crashes? This is a question no one can answer right now.

A more fundamental shortcoming is standards. The billing scopes and statistical methods for tokens on different large model platforms are not unified, and a third-party neutral audit system is currently still a void. Xie Baojian, Vice Dean of the Institute of Southern Advanced Finance at Jinan University, pointed out the crux: from benchmark demonstrations to large-scale replication, what’s holding it back isn’t willingness, but standards.

Banks’ responses are also on the table: not taking tokens as the sole basis, but “contract + token call data + transaction flows” for three-way reconciliation, verifying the entire process of token production, calling, and settlement; and striving to get computing power platforms to open system query interfaces to banks for real-time monitoring; companies already granted credit in Guangzhou are mainly based on credit loans, while also being included in the municipal government-bank risk-sharing mechanism. The official tone of Haizhu District is also very restrained: token consumption volume serves only as one of the reference standards for credit granting, not directly equivalent to collateral itself.

The Real Takeaways of This Matter

Following this news, I suggest keeping an eye on three questions.

First, is this “rights confirmation” or “hype rhetoric”? The biggest gift “Token Loans” give to AI enterprises might not be the few million in loans, but an official acknowledgment: continuously and stably consuming tokens is equivalent to self-proving to the financial system “I have real business and stable customers.” When “consumption volume” becomes the language of credit, companies that get loans will be more motivated to run their business on platforms that are measurable and auditable—this is good for computing power platforms and local governments. But we must be wary of the other side: once “high consumption” itself becomes a financing asset, distorted demands for burning money just for the sake of loans will appear. To judge if such products are healthy, just盯住 one indicator: does the money lent out match the real business returns?

Second, who is bearing the risk now? Currently, the structure everywhere is generally “credit loan + government-bank risk sharing”—government subsidies and risk compensation guarantees, so banks dare to lend money to companies without collateral. This is a reasonable trial cost, but it also means that risk pricing at this stage is largely determined by policy, not the market. The observation window is the moment when the risk-sharing mechanism expires or expands: if banks are still lending then, it means data risk control has really stood up; if the tide recedes along with subsidies, then this round of “Token Loans” is just a phased industrial stimulus.

Third, will this ruler measure beyond the AI industry? The logic of token usage volume—using machine-generated operational traces instead of reports—theoretically applies to all digitally operated industries: e-commerce, short dramas, games, software services. Once it runs smoothly, it may rewrite not just the financing of AI companies, but the evaluation method of credit for SMEs as a whole. Of course, the prerequisite is to fill the gaps in standards, auditing, and anti-brushing. The title of that People’s Daily investigation asks what is “loaned”—for now, what is loaned is a whole new set of “credit based on usage” gameplay, and the old question this new gameplay must answer: the data is real, but where does the money for repayment come from?

References

People’s Daily — People’s Daily “What Do ‘Token Loans’ Bring to AI Enterprises?” (Finance In-Depth, September 14, 2026, Page 18, Reporter Cheng Yuanzhou)
Economic Daily “Token Loans Alleviate AI Enterprise Financing Dilemmas” (August 31, 2026, Reporter Yu Jian; Reproduced in full by Xinhuanet)
The Beijing News “First Batch of Credit Approvals Near 2 Billion! Beijing E-Town Takes the Lead in Landing Token Computing Power Loans Citywide” (September 13, 2026, Source “Beijing E-Town Innovation Release”)
IT Home citing CCTV Finance: CIFTIS “Finance + Sci-Tech” Joint Exhibition, Computing Power Token Loans Accelerate Promotion After Landing in August (September 12, 2026); Beijing E-Town “Yiqi Token Loan” Lands (September 5, 2026, Source “Beijing E-Town” Official Account)
Cailian Press: Chongqing Supports Financial Institutions in Exploring New Models Like “Computing Power Banks” and “Token Vouchers + Token Loans” (September 9, 2026, Title-level verification)
Red Star News “Red Star Observation”: From “Subsidizing Computing Power” to “Subsidizing Tokens”, Chengdu’s 100 Million Token Vouchers (September 12, 2026, Title-level verification)